On the Least Median Square Problem
|
|
- Loren Warren
- 6 years ago
- Views:
Transcription
1 On the Least Median Square Problem Jeff Erickson Sariel Har-Peled Department of Computer Science University of Illinois at Urbana-Champaign {jeffe,sariel} David M. Mount Department of Computer Science and Institute for Advanced Computer Studies University of Maryland mount ABSTRACT We consider the exact and approximate computational complexity of the multivariate LMS linear regression estimator. The LMS estimator is among the most widely used robust linear statistical estimators. Given a set of n points in IR d and a parameter k, the problem is equivalent to computing the narrowest slab bounded by two parallel hyperplanes that contains k of the points. We present algorithms for the exact and approximate versions of the multivariate LMS problem. We also provide nearly matching lower bounds for these problems, under the assumption that deciding whether n given points in IR d are affinely nondegenerate requires Ω(n d ) time. Categories and Subject Descriptors: F.2.2 [Analysis of algorithms and problem complexity]: Nonnumerical problems and algorithms Geometric problems and algorithms; G.3 [Probability and statistics]: Robust regression Keywords: LMS regression, robust statistics, approximation algorithms, lower bounds General Terms: Algorithms, Theory 1. INTRODUCTION Fitting a hyperplane to a finite collection of points in space is a fundamental problem in statistical estimation. Robust estimators are of particular interest because of their insensitivity to outlying data. The principal measure of the robustness of an estimator is its breakdown value, that is, the fraction (up to 50%) of outlying data points that can corrupt See jeffe/pubs/halfslab.html for the most recent version of this paper. Partially supported by NSF CAREER award CCR and NSF ITR grants DMR and CCR Partially supported by NSF CAREER award CCR Partially supported by NSF grant CCR Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SCG 04, June 8 11, 2004, Brooklyn, New York, USA. Copyright 2004 ACM /04/ $5.00. the estimator. Rousseeuw s least median-of-squares (LMS) linear regression estimator [22] is among the best known and most widely used robust estimators. The LMS estimator (with intercept) is defined formally as follows. Consider a set P = {p 1, p 2,..., p n} of n points in IR d, where p i = (x i,1, x i,2,..., x i,d 1, x i,d ). We would like to compute a parameter vector θ = (θ 1, θ 2,..., θ d ) that best fits the data by the linear model x i,d = x i,1θ 1 + x i,2θ x i,d 1 θ d 1 + θ d + r i for all i, where r 1, r 2,..., r n are the (unknown) errors, or residuals. The LMS estimator is defined to be the parameter vector θ that minimizes the median of the squared residuals. More generally, given a parameter k, where d + 1 k n, the problem is to find a parameter vector θ that minimizes the kth smallest squared residual. This more general problem is also called the least-quantile squared (LQS) estimator [24]. Typically, we are interested in the case where both k = Ω(n) and n k = Ω(n). This estimator is widely used in practice, for example in finance, chemistry, electrical engineering, process control, and computer vision [23]. In addition to having a high breakdown value, the LMS estimator is regression-, scale-, and affine-equivariant, which means that the estimate transforms properly under these types of transformations [24]. The LMS estimator may be used either alone or as an initial step in more complex estimation schemes [26]. The LMS estimator has an easy geometric interpretation. A slab is the volume enclosed by two parallel hyperplanes; the height of a slab is the vertical separation between its bounding hyperplanes. Computing the LMS estimator is clearly equivalent to computing a slab of minimum height that encloses at least k of the points. The LMS estimator is obtained as the hyperplane that bisects this slab. In the dual setting, the problem is equivalent to finding the shortest vertical segment in IR d that intersects at least k of a set of n given hyperplanes. We also consider the closely related problem of computing the slab of minimum width enclosing k points, where the width of a slab is measured normal to its bounding hyperplanes. This problem is also called LMS orthogonal regression [24]. The most efficient exact algorithm for computing the LMS estimator in the plane is a topological plane-sweep algorithm of Edelsbrunner and Souvaine [7], which runs in O(n 2 ) time and requires O(n) space. In higher dimensions, an O(n d+1 log n)-time algorithm has long been known [19, 24]. The LMS problem can also be expressed as a linear programming problem with violations; using this formulation,
2 the problem can be solved in O(n(n k) d+1 ) time [18, 4]. An alternative approach that can also be deployed in this case is an algorithm by Har-Peled and Wang for approximate shape fitting with outliers [16], which extracts a small coreset and finds the best possible solution within this coreset. However, this algorithm excels only when the number of outliers is small; specifically, the running time is near-linear only if k n o(n 1/2d ). The case of most interest to us, where n k = Ω(n), is the worst case for these algorithms. It is natural to ask whether either of these approaches can be extended to handle the case where the number of outliers is large. Finally, the problem of fitting the data with two slabs of minimum width seems to be inherently connected to the problem of LMS; intuitively, we want to find a good clustering for most of the points by the first slab, and also a good clustering for the remaining (outlier) points. Thus, a better understanding of the LMS problem might lead to a better understanding of the two-slab problem, which seems to be surprisingly hard [15]. Currently no near-linear-time approximation algorithm is known for d > 3. Thus, LMS is a fundamental problem, and a better understanding of it would lead to a better understanding of several central optimization problems. Given the high complexity of computing the LMS estimator exactly, it is natural to consider whether more efficient approximation algorithms exist. Mount et al. [20] presented a practical approximation algorithm for the LMS line estimator in the plane, based on approximating the quantile and/or the vertical width of the slab. Their algorithm, however, does not guarantee a better than O(n 2 ) running time when the quantile is required to be exact. Olson presented a 2-approximation algorithms for LMS, which runs in O(n log 2 n) time in the plane and in O(n d 1 log n) time for any fixed d 3 [21]. In this paper we consider the computational complexity of the both the exact and approximate versions of the LMS problem. In Section 3, we describe a randomized algorithm to compute the exact LMS estimator for n points in IR d in O(n d log n) time with high probability. We also describe, in Section 4, a randomized ε-approximation algorithm whose running time is O((n d /kε) polylog) with high probability. For the most interesting case k = Ω(n), this is faster than our exact algorithm by roughly a factor of n/ε. In Section 5, we provide results suggesting that any exact algorithm for the LMS problem requires Ω(n d ) time, and any constant-factor approximation algorithm requires Ω(n d 1 ) time, thus providing a strong indication that our algorithms are close to optimal. Specifically, we describe linear-time reductions from the affine degeneracy problem: Given a set of n points on the d-dimensional integer 1 lattice Z d, do any d + 1 of the points lie on a common hyperplane? These reductions imply lower bounds if the following widely-believed conjecture is true. Conjecture 1.1 Solving the d-dimensional affine degeneracy problem requires Ω(n d ) time in the worst case. If this conjecture is true, the Ω(n d ) lower bound is tight; the problem can be solved in O(n d ) time by constructing 1 In the algebraic decision tree model of computation, the restriction to integers can be removed by using formal infinitesimals in our reductions [10, 11]. the dual hyperplane arrangement. Erickson and Seidel [13, 12] proved an Ω(n d ) lower bound on the number of sidedness queries required to solve this problem; however, the model of computation in which their lower bound holds is not strong enough to solve the LMS problem, since it does not allow us to compare widths of different slabs. The strongest lower bound known in any general model of computation is Ω(n log n), for any fixed dimension, in the algebraic decision and computation tree models [2, 25], although the problem is known to be NP-complete when d is not fixed [17, 12]. The planar affine degeneracy problem is one of Gajentaan and Overmars canonical 3sum-hard problems [14]. (In higher dimensions, the affine degeneracy problem is actually (d + 1)-sum-hard [12]; however, the best lower bound that this could imply is only Ω(n d/2 +1 ) [11].) 2. NOTATION AND TERMINOLOGY Whenever we work in the space IR d, we will refer to the x d -coordinate direction as vertical and each of the other coordinate directions as horizontal. A hyperplane is vertical if it contains a vertical line and horizontal if its normal direction is vertical. A slab is the non-empty intersection of two closed halfspaces whose bounding hyperplanes are parallel. We will distinguish between two different natural notions of the thickness of a slab. The height of a slab σ, denoted ht(σ), is the length of a vertical line segment with one endpoint on each bounding hyperplane. If one slab has smaller height than another, we say that the first slab is shorter and the second is taller. We denote by ht(p, k) the height of the shortest slab containing at least k points from a point set P. On the other hand, the width of a slab σ, denoted wd(σ), is the distance between the two bounding hyperplanes, measured along their common normal direction. If one slab has smaller width than another, we say that the first slab is narrower and the second is wider. We denote by wd(p, k) the width of the narrowest slab containing at least k points in a point set P. A slab whose height or width is zero is just a hyperplane. Vertical slabs, even those with zero width, have infinite height. The LMS problem has a natural dual formulation, which is crucial in the development of our algorithms. We use the standard duality transformation that maps any point (a 1, a 2,..., a d ) to the hyperplane x d = a 1x 1 + a 2x a d 1 x d 1 x d and vice versa. Under this duality map, a point p is above (resp. below) a hyperplane h if and only if the dual hyperplane p is below (resp. above) the dual point h ; moreover, the vertical distance between the point and the hyperplane is preserved. The dual of a slab σ is a vertical line segment σ of length ht(σ); a point p lies inside the slab σ if and only if the dual hyperplane p intersects the dual segment σ. Thus, the LMS problem is equivalent to finding, given a set H of hyperplanes, the shortest vertical segment that intersects at least k of them. To be consistent with the primal formulation, we let ht(h,k) denote the length of this segment. For any hyperplane h and any real value t, let h + t denote the hyperplane resulting from translating h upward by (vertical) distance t, and let σ(h, t) denote the slab bounded by h and h + t.
3 3. EXACT ALGORITHMS 3.1 Minimum Height In this section, we develop efficient exact algorithms for computing the LMS estimator in any fixed dimension. The two-dimensional problem can be solved in O(n 2 ) time using Edelsbrunner and Souvaine s dual topological sweep algorithm [7]. Our higher-dimensional algorithm improves the previous best running time by a factor of n. Theorem 3.1 Given a set H of n hyperplanes in IR d and an integer k, we can compute the shortest vertical segment that stabs k hyperplanes of H in O(n d log n) time with high probability. Proof: Clearly, the endpoints of the required segment lie on at least d + 1 hyperplanes of H. Let T be the (multi)set of critical values t such that the arrangement of slabs Σ(t) = {σ(h, t) h H} contains a vertex on d+1 bounding hyperplanes. For notational convenience, let t = ht(h,k). We easily observe that T = O(n d+1 ) and that t T. A vertical line segment of length t stabs k hyperplanes in H if and only if some point lies in at least k slabs in Σ(t), or equivalently, if and only if some vertex has depth at least k in the arrangement of Σ(t). Thus, given a candidate value t, we can decide whether t < t in time O(n d ) by constructing the arrangement of Σ(t) and computing the depth of every vertex. We transform this decision procedure into a search algorithm as follows. We randomly choose a sample R of O(n d ) vertical distances from T, by repeatedly choosing d + 1 hyperplanes of H and computing the shortest vertical segment stabbing all of them. We then use the decision procedure to guide a binary search through the random sample R. In O(n d log n) time, this binary search finds an interval [t, t + ] that contains t but no other elements of R. With high probability, the interval [t, t + ] contains only O(n log n) values from T. To locate t within this interval, we maintain the arrangement of slabs Σ(t) in a kinetic data structure [1] as t varies continuously from t = t to t = t +. The combinatorics of the arrangement changes only at times in T. At each such critical event, a simplicial cell in the arrangement of Σ(t) collapses to a point and reverses orientation. We maintain a priority queue of collapse times for the simplicial cells in Σ(t), ignoring any cell that collapses after time t +. We prepare the initial priority queue by constructing the arrangement of Σ(t ), computing the collapse time of each simplicial cell (in O(1) time), and inserting only the relevant future events into the queue. Each critical event requires a constant number of priority queue operations, plus O(1) additional work. Thus, the total cost of the sweep is O(n d + E log E), where E is the number of critical events. With high probability, E = O(n log n), so the sweep requires only O(n d ) time. Corollary 3.2 Given a set P of n points in IR d and an integer k, we can compute the shortest slab containing k points of P in O(n d log n) time with high probability. Chan [5] observes that his randomized optimization technique [3] can be used in place of sampling and sweeping, using a constant-complexity cutting to generate subproblems, as in his other results on linear programming with violations [4]. Chan s techniques reduce the expected running time of our algorithm to O(n d ) but do not improve our high-probability time bound. 3.2 Minimum Width We can use similar techniques to solve the orthogonal LMS regression problem. A simple modification of Edelsbrunner and Souvaine s topological sweep algorithm [7] solves the two-dimensional version of this problem in O(n 2 ) time and O(n) space. Theorem 3.3 Given a set P of n points in IR d and an integer k, we can compute the narrowest slab containing k points of P in O(n d log n) time with high probability. Proof: We easily observe that the target slab has at least one point (in fact, d + 1 points) from P on its boundary. We describe an algorithm to find the narrowest slab that contains k points in P and has a specific point q P on its boundary. To find the narrowest unrestricted slab containing k points in P, we simply run this algorithm once for every reference points q P and return the narrowest result. We first solve the following decision problem: Is there a slab of width w, whose boundary contains q, that contains k points in P? Any slab with q on its boundary is uniquely determined by the normal vector pointing into the slab from the bounding hyperplane that contains q. Thus, we have a duality transformation from slabs with q on their boundary and points on the sphere of directions S d 1 centered at q. For any point p P, the family of slabs of width w that contain p defines a strip s(p,w) on S d 1 ; this strip is the intersection of S d 1 and a slab with q on its boundary. Thus, our decision problem can be rephrased as follows: Is any point in S d 1 covered by k of the strips in S(w) = {s(p, w) p P }? This problem can be decided in O(n d 1 + nlog n) time by computing the arrangement of strips S(w). We can now transform this decision procedure into a search algorithm, increasing the running time by a factor of O(log n), using the same random sampling and sweeping techniques used to prove Theorem 3.1. We omit the remaining straightforward details. Again, Chan s randomized optimization technique [3, 4] can be used to reduce the expected running time of our algorithm to O(n d ). 4. APPROXIMATION ALGORITHMS Our algorithm for approximating the LMS estimator rely on the same techniques used in our exact algorithms. Specifically, we exploit the dual formulation of the problem, we use random sampling and sweeping to reduce the search to a decision problem, and we begin by considering only slabs with a given reference point (or dual hyperplane). Lemma 4.1 Give a set H of n hyperplanes in IR d, a hyperplane g in IR d (not necessarily in H), and an integer k, we can compute the shortest vertical segment that stabs k hyperplanes in H and whose midpoint lies on g, in O(n d 1 log n+ nlog 2 n) time with high probability.
4 Proof: As usual, we reduce the search problem to a decision problem: Given a parameter l 0, is there a vertical segment of length l, whose midpoint lies on g, that stabs k hyperplanes of H? For every hyperplane h H, let τ(h, l) be the set of points on g that lie at vertical distance at most l/2 from h, or equivalently, the intersection of g with the slab of height l centered at h. Generically, τ(h, l) is a (d 1)-dimensional slab in g; however, if h is parallel to g, then τ(h, l) is either the empty set or all of g. Finally, let T(l) = {τ(h, l) h H}. Our decision problem can now be rephrased as follows: Is any point in g covered by k or more regions in T(l)? This problem can be decided in O(n d 1 + nlog n) time by simply constructing the arrangement of T(l) and computing the depth of every vertex. To complete the proof of the theorem, we apply the usual random sampling and sweeping techniques to our decision procedure. If d > 2, applying this lemma once for every hyperplane in g H gives us an alternate proof of Theorem 3.1. Once again, we can save a factor of O(log n) in the expected running time using Chan s randomized optimization technique [3, 4]. Theorem 4.2 Given a set H of n hyperplanes in IR d, an integer k, and a real parameter ε > 0, we can compute a vertical segment of length at most (1 + ε) ht(h,k) that stabs k hyperplanes of H in O((n d /kε) polylog n) time with high probability. Proof: Let s be the shortest vertical segment that stabs k hyperplanes in H. Let R be a random sample of O((n/k) log n) hyperplanes in H. By ε-net theory, s stabs at least one of the hyperplanes of R with high probability. For every hyperplane in R, apply the algorithm of Lemma 4.1, and let s be the shortest segment thus computed. Clearly, the length of s is at most twice the length of s. Now let δ = ε s /4, and consider the set of hyperplanes R = {h + iδ h R, i = 4/ε,..., 4/ε }, where as usual h+t is the hyperplane resulting by translating h vertically by distance t. With high probability, at least one of the hyperplanes of R intersects s and is within distance δ = ε s /4 < ε s /2 = ε ht(h,k)/2 from the midpoint of s. Thus, the shortest vertical segment computed by applying the algorithm of Lemma 4.1 to each hyperplane of R is the required (1 + ε)-approximation. The algorithm of Lemma 4.1 is invoked O((n/kε) log n) times, so the overall running time is O((n 2 /kε) log 3 n) if d = 2 and O((n d /kε)log 2 n) otherwise. The algorithm of Theorem 4.2 clearly also solves the dual problem of finding an approximately shortest slab containing k points. Corollary 4.3 Given a set P of n points in IR d, an integer k, and a real parameter ε > 0, we can compute a slab of height at most (1 + ε) ht(p, k) that contains k points of P in O((n d /kε) polylog n) time with high probability. In the special case k = Ω(n), we can improve the running time of our algorithm by a factor of O(log n) using Chazelle s deterministic algorithm for constructing ε-nets [6]. Finding an approximately narrowest slab is only slightly different. Using a similar algorithm to the one described above, together with the techniques of Theorem 3.3, we obtain the following result. We omit further details. Theorem 4.4 Given a set P of n points in IR d, an integer k, and a real parameter ε > 0, we can compute a slab of width at most (1 + ε)wd(p, k) that contains k points of P in O((n d /kε) polylog n) time with high probability. 5. HARDNESS RESULTS In this section, we prove relative lower bounds for both the exact and approximate LMS hyperplane-fitting problems. Our reductions suggest that computing the exact LMS estimator requires Ω(n d ) time and that computing any constant-factor approximation requires Ω(n d 1 ) time, in the worst case. Thus, it is unlikely that our LMS algorithms can be sped up by more than polylogarithmic factors, at least when k and n k are both Ω(n). Our reductions also directly imply Ω(n log n) lower bounds for both the exact and approximate LMS problems in any fixed dimension, in the algebraic decision tree model; they also imply that when the dimension is not fixed, even the approximate LMS problem is NP-hard. Our reductions rely on the following observations. We say that a slab is minimal for a set of points if its boundary contains at least d + 1 affinely independent points from the set. For any set P of more than d points, the shortest and narrowest slabs containing P are both minimal. Lemma 5.1 Let P be a set of d + 1 affinely independent points on the integer grid [ M.. M] d, and let σ be any minimal slab containing P. (a) ht(σ) can be written as a ratio of integers p/q, where p and q are both O(M d ). (b) Either ht(σ) = 0 or ht(σ) = Ω(1/M d ). (c) Either wd(σ) = 0 or wd(σ) = Ω(1/M d ). (d) σ has an integer normal vector n(σ) whose coefficients have absolute value O(M d ). (e) wd(δ) 2 can be written as a ratio of integers p /q, where p and q are both O(M 4d ). (All these bounds hide constant factors exponential in d.) Proof: Without loss of generality, we assume the slab σ is not vertical. We refer to d-dimensional Lebesgue measure as volume and (d 1)-dimensional Lebesgue measure as area. (a) Let denote the convex hull of P and let V denote the volume of. We can express V as the ratio A ht(σ)/d, where A is the sum of the signed areas of the vertical projections of certain facets of onto IR d 1. Specifically, a facet contributes its area to A if it touches the lower bounding hyperplane of σ, positively if is locally above that facet and negatively otherwise. Clearly, V = O(m d ), and the projected area of each
5 facet is at most O(M d 1 ). Moreover, since every coordinate is an integer, V is an integer multiple of 1/d!, and the projected area of each facet is an integer multiple of 1/(d 1)!. We conclude that ht(σ) = dv A = d! V (d 1)! A, where the integers d! V and (d 1)!A are both O(M d ). (b) This follows immediately from part (a). (c) Consider a line through the origin normal to the hyperplanes bounding σ. This line forms an angle of at most 45 with at least one coordinate axis. If necessary, reflect P and σ across some hyperplane x i = x d so that that axis is vertical. We now have wd(σ) ht(σ)/ 2, and the result follows from part (b). (d) Write P = {p 0, p 1,..., p k, q k+1,..., q d }, where each point p i lies on the lower bounding hyperplane of σ and each point q i lies on the upper bounding hyperplane of σ. We define d 1 integer vectors v 1,..., v d 1 parallel to σ as follows: if i k, we take v i = p i p i 1, and if i > k, we take v i = q i q i+1. These vectors are linearly independent, since otherwise P would be affinely degenerate. The exterior product v 1 v 2 v d is a vector normal to σ. Each component of this exterior product is the determinant of a (d 1) (d 1) minor of the (d 1) d matrix of coordinates v ij. Since each of these coordinates is an integer with absolute value O(M), each component of the normal vector is an integer with absolute value O(M d 1 ). (e) This follows immediately from the identity wd(σ) = ht(σ)n d (σ)/ n(σ), where n(σ) is the vector normal to σ constructed in part (d), and n d (σ) is its vertical component. 5.1 Exact Height We now establish our (relative) lower bounds for computing minimum-height slabs exactly. Theorem 5.2 Conjecture 1.1 implies that computing the shortest slab containing d + 1 points from a given set of n points in Z d requires Ω(n d ) time in the worst case. Proof: The given points are affinely degenerate if and only if the shortest slab containing d+1 points has height zero. Theorem 5.3 Conjecture 1.1 implies that computing the shortest slab containing n/2 points from a given set of n points in Z d requires Ω(n d ) time in the worst case. Proof: Suppose we are given a set P of m = n/2 d 1 points in Z d. Let x + d and x d be the largest and smallest x d- coordinates of any point in P, respectively. In O(n) time, we can construct a new set Q of 2m + 2(d + 1) = n points by taking the union of P, a copy of P shifted upward by 2(x + d x d ), and a set of n 2m = 2(d + 1) extra points at least 5(x + d x d ) above everything else. The original set P contains d + 1 points on a common hyperplane if and only if the shortest slab that contains n/2 points in Q has height exactly 2(x + d x d ). This reduction can be generalized easily to either larger or smaller numbers of points in the target slab, as follows: Theorem 5.4 Conjecture 1.1 implies that computing the shortest slab containing k points from a given set of n points in Z d requires Ω(min{k, n k} d ) time in the worst case. Proof: If k < n/2, we start with a set P of m = k d 1 points. We construct a new set Q containing two copies of P, one directly above the other, with n 2m extra points far above both copies. Similarly, if k > n/2, we start with a set P of m = n k + d + 1 points. We construct a new set Q contains two copies of P, one directly above the other, with n 2m extra points directly between the two copies. In both cases, the shortest slab containing k points of Q has height equal to the vertical distance between the two copies of P if and only if P is affinely degenerate. Thus, Conjecture 1.1 implies that computing the shortest slab containing k points in Q requires Ω(m d ) time. In the first case, the extra points lie above the shortest slab; in the second case, the extra points are inside the shortest slab. Theorem 5.5 Conjecture 1.1 implies that computing the shortest slab containing k points from a given set of n points in Z d requires Ω((n/k) d ) time in the worst case. Proof: Suppose we are given a set P of m = n(d + 1)/k points in Z d. In linear time, we compute an upper bound M on the absolute value of every coordinate. Choose an appropriate constant δ = O(1/kM d+1 ). We construct a new set P consisting of k/(d+1) copies of P, where the ith copy is shifted upward a distance of iδ. If some hyperplane h contains d + 1 points in P, then the slab σ(h,(k/(d + 1) 1)δ) contains k points in P, and thus ht(p, k) (k/(d + 1) 1)δ = O(1/M d+1 ). On the other hand, suppose P is affinely nondegenerate. Let σ be any slab containing k points in P. This slab contains copies of at least d + 1 points in P, so its height is at least ht(p, d + 1) (k/(d + 1) 1)δ = Ω(1/M d ) by Lemma 5.1. These reductions suggest Ω(n d ) lower bounds for exact LMS for the most interesting case, when both k and n k are Ω(n), as well as the simplest nontrivial case k = O(1). We conjecture that the true complexity is Ω((n k) d ) for any k, but there is still a gap between the upper and lower bounds for most small values of k. Since the affine degeneracy problem is NP-hard when the dimension d is not fixed [17, 12], our reductions imply a similar NP-hardness result for the LMS estimator. Corollary 5.6 For any k n Ω(n c ) for some constant c, computing the shortest slab containing k points from a given set of n points in Z n is NP-hard. 5.2 Approximate Height Theorem 5.7 Conjecture 1.1 implies that computing a slab of height at most 2ht(P, k) containing k points from a given set P of n points in Z d requires Ω((n k) d 1 ) time in the worst case.
6 Proof: Suppose we are given a set P of m = n/2 k/2 d 1 points on the integer lattice Z d 1. Let M be an upper bound the maximum absolute value of any coordinates in P, and let δ = 1/(d 2)!(2M) d 2 ; we can compute these values in O(m) time. In O(n) time, we construct a new set Q comprised of three subsets: (1) a copy of P on the vertical hyperplane x 1 = 1, (2) a set of k 2(d + 1) points within distance δ/10 of the origin, all on the hyperplane x 1 = 0, and (3) a copy of S (the reflection of P through the origin) on the hyperplane x 1 = 1. For any non-vertical slab σ, let σ x denote the intersection of σ with the hyperplane x 1 = x; this is a (d 1)-dimensional slab with the same height as σ. If any d points of P lie on a common (d 2)-flat, then there is a slab of height at most δ/5 containing k points of Q. Otherwise, let σ be any slab containing k points of Q. Without loss of generality, σ 1 contains at least d + 1 points of P, so by Lemma 5.1(1), we have ht(σ) = ht(σ 1) δ. Thus, by approximating H k (S ) within a factor of 2, we can determine whether the original set P contains a degeneracy. Conjecture 1.1 implies that this requires Ω(m d 1 ) = Ω((n k) d 1 ) time in the worst case. The proof of Theorem 5.5 implies an even stronger hardness result when k is small. Theorem 5.8 Conjecture 1.1 implies that for any function f(n, k), computing a slab of height at most f(n, k) ht(p, k) containing k points from a given set P of n points in Z d requires Ω((n/k) d ) time in the worst case. Proof: Suppose we are given a set P of m = n(d + 1)/k points in the finite integer lattice [ M.. M] d. Following the proof of Theorem 5.5, we can construct a set P of n points in O(n) time, such that approximating ht(p, k) up to a factor of M tells us whether P is affinely degenerate. We can replace this factor M by any function f(n, k) by setting δ = 1/kM d f(n, k) d in the reduction. Our reductions strongly suggest that our approximation algorithm is within a polylogarithmic factor of optimal, at least when k = Ω(n) or k = O(1). They also imply that the approximate LMS problem is NP-hard when the dimension is not fixed. Corollary 5.9 For any k n Ω(n c ) for some constant c, computing a 2-approximation of the shortest slab containing k points from a given set of n points in Z n is NP-hard. 5.3 Reducing Height to Width Finally, we describe a general reduction from computing slabs with minimum height to computing slabs of minimum width. This reduction implies that all our lower bounds for minimizing height apply verbatim to the corresponding width problem. The key observation is that horizontally scaling Z d does not change the height of any slab, although it does change the width. If we scale any point set P far enough, then sorting the non-vertical minimal slabs by width would be the same as sorting them by height; in particular, the narrowest non-vertical slab containing k points of P will also be the shortest slab containing k points of P. There are two main technical difficulties: quantifying the amount of scaling required and eliminating vertical slabs from consideration. Suppose we want to find the shortest slab containing k d + 1 points from a given set P of n points on the integer lattice [ M.. M] d. If M is not given, we can easily compute it in O(n) time. Let P be the set obtained by scaling P horizontally (that is, in every direction except vertically) by a large integer factor := Ω(M 6d ). Scaling any slab σ horizontally by gives us a slab σ with the same height, containing the corresponding subset of points. Fix a minimal non-vertical slab σ containing at least d+1 points of P. Let n be the integer normal vector of σ described by Lemma 5.1(c). To obtain a normal vector n for the scaled slab σ, we can simply scale n in the vertical direction by a factor of. We can decompose n into a vertical component n v and a horizontal component n h. Lemma 5.1(c) implies that n h = O(M d ), and since σ is not vertical, n v = Ω(M 6d ). We have the following bound on the width of σ in terms of its height: p n ht(σ ) = wd(σ v 2 + n h ) 2 n v s wd(σ 1 ) 1 + = wd(σ ) 1 + = wd(σ ) + Ω(M 5d ) «1 Ω(M 5d ) 1 Ω(M 3d ) Lemma 5.1(a) implies that the heights of any two minimal slabs σ 1 and σ 2 either are equal or differ by at least Ω(1/M 2d ). It follows that ht(σ 1) < ht(σ 2) implies wd(σ 1) < wd(σ 2); the height order and width order of the non-vertical minimal slabs is the same, except that some equal-height pairs may not have equal width. In particular, the narrowest non-vertical slab containing k points in P is also the shortest such slab. The entire reduction requires only linear time, and increases the bit length of the input by at most a factor of O(d). This completes the reduction from finding the shortest slab containing k of n given points to finding the narrowest such slab that is not vertical, but what about vertical slabs? Lemma 5.1(c) implies that the narrowest vertical slab σ v containing k points of P is either a single hyperplane or it has width Ω( /M d 1 ) = Ω(M 5d ). Thus, if no vertical hyperplane contains k points in P, the narrowest slab containing k points in P is not vertical, since the entire point set fits in a slab of width 2M, so our reduction is complete. However, if k points in P lie on a vertical hyperplane, the shortest and narrowest slabs containing k points in P may not coincide. To avoid this problem, we first perturb the initial set P, essentially following the infinitesimal perturbation method of Emiris and Canny [8, 9]. Let δ = 1/M 4d. For any point p P, let p denote a point at distance at most δ from P, and let P = { p p P }. For any minimal slab σ containing some subset of P, we define a corresponding slab σ that is minimal for the corresponding subset of P. Lemma 5.1 implies that ht(σ) O(M d )wd(σ), so ht( σ) ht(σ) + O(1/M 3d ), and that two minimal slabs for P differ in height by at least 1/M 2d. It follows that if σ 1 is shorter than σ 2, then σ 1 is shorter than σ 2. In other words, the shortest slab containing
7 k points in P has the same combinatorial description as some shortest slab containing k points in P. Arbitrarily index the points in P as p 1, p 2,..., p n, and let q be the smallest prime number larger than n (and therefore less than 2n). We choose the specific perturbation p i = p i + δµ q(i), where µ q is the modular moment curve µ q(t) := 1 q (t, t2 mod q, t 3 mod q,..., t d mod q). We can express the volume of any simplex in P as a polynomial in δ. Lemma 5.1 implies that the sign of this polynomial is determined by the sign of the largest term. Moreover, the coefficient δ d term is the volume of a simplex whose vertices are integer points on the modular moment curve, and is therefore not equal to zero. We conclude that no d + 1 points in P lie on a common hyperplane; in particular, no k points lie on a vertical hyperplane. Scaling the set P by a factor of q/δ gives us an integer point set, where every coordinate has absolute value at most O(M 5d ). Thus, to find the shortest slab containing k points in P, we can apply our earlier reduction to P. The entire reduction requires only linear time and increases the bit length of the input by at most a factor of O(d 2 ). Theorem 5.10 Conjecture 1.1 implies that computing the narrowest slab containing k points from a given set of n points in Z d requires time Ω(min{k, n k} d ) and Ω((n/k) d ) in the worst case. Theorem 5.11 Conjecture 1.1 implies that computing a slab of width at most 2 wd(p, k) containing k points from a given set P of n points in Z d requires time Ω((n k) d 1 ) and Ω((n/k) d ) in the worst case. Corollary 5.12 For any k n Ω(n c ) for some constant c, computing a 2-approximation of the thinnest slab containing k points from a given set of n points in Z n is NP-hard. Acknowledgments We thank Timothy Chan for suggesting the use of his randomized optimization technique. References [1] J. Basch, L. J. Guibas, and J. Hershberger. Data structures for mobile data. Proc. 8th ACM-SIAM Sympos. Discrete Algorithms, , [2] M. Ben-Or. Lower bounds for algebraic computation trees. Proc. 15th Annu. ACM Sympos. Theory Comput., 80 86, [3] T. M. Chan. Geometric applications of a randomized optimization technique. Discrete Comput. Geom. 22: , [4] T. M. Chan. Low-dimensional linear programming with violations. Proc. 43th Annu. IEEE Sympos. Found. Comput. Sci., , [5] T. M. Chan. Personal communication, Feb [6] B. Chazelle. The Discrepancy Method. Cambridge University Press, [7] H. Edelsbrunner and D. L. Souvaine. Computing median-of-squares regression lines and guided topological sweep. J. Amer. Statist. Assoc. 85: , [8] I. Emiris and J. Canny. A general approach to removing degeneracies. SIAM J. Comput. 24: , [9] I. Z. Emiris, J. F. Canny, and R. Seidel. Efficient perturbations for handling geometric degeneracies. Algorithmica 19(1 2): , [10] J. Erickson. On the relative complexities of some geometric problems. Proc. 7th Canad. Conf. Comput. Geom., 85 90, [11] J. Erickson. Lower bounds for linear satisfiability problems. Chicago J. Comput. Sci. 1999(8), [12] J. Erickson. New lower bounds for convex hull problems in odd dimensions. SIAM J. Comput. 28: , [13] J. Erickson and R. Seidel. Better lower bounds on detecting affine and spherical degeneracies. Discrete Comput. Geom. 13:41 57, [14] A. Gajentaan and M. H. Overmars. On a class of O(n 2 ) problems in computational geometry. Comput. Geom. Theory Appl. 5: , [15] S. Har-Peled. On core-sets and slabs. Manuscript, [16] S. Har-Peled and Y. Wang. Shape fitting with outliers. Proc. 19th Annu. ACM Sympos. Comput. Geom., 29 38, [17] L. Khachiyan. On the complexity of approximating determinants in matrices. J. Complexity 11: , [18] J. Matoušek. On geometric optimization with few violated constraints. Discrete Comput. Geom. 14: , [19] P. Meer, D. Mintz, A. Rosenfeld, and D. Y. Kim. Robust regression methods for computer vision: A review. Int. J. Computer Vision 6:59 70, [20] D. M. Mount, N. S. Netanyahu, K. R. Romanik, R. Silverman, and A. Y. Yu. A practical approximation algorithm for the LMS line estimator. Proc. 8th ACM-SIAM Sympos. Discrete Algorithms, , [21] C. F. Olson. An approximation algorithm for least median of squares regression. Inform. Process. Lett. 63: , [22] P. J. Rousseeuw. Least median of squares regression. J. Amer. Statist. Assoc. 79(388): , [23] P. J. Rousseeuw. Introduction to positive breakdown methods. Handbook of Statistics, Vol. 15: Robust Inference, , Elsevier. [24] P. J. Rousseeuw and A. M. Leroy. Robust Regression and Outlier Detection. John Wiley & Sons, New York, [25] A. C. Yao. Lower bounds for algebraic computation trees with integer inputs. SIAM J. Comput. 20(4): , [26] V. J. Yohai. High breakdown-point and high efficiency robust estimates for regression. Ann. Stat. 15: , 1987.
On the Least Median Square Problem
On the Least Median Square Problem Jeff Erickson Sariel Har-Peled David M. Mount July 18, 2005 Abstract We consider the exact and approximate computational complexity of the multivariate LMS linear regression
More informationLine Arrangement. Chapter 6
Line Arrangement Chapter 6 Line Arrangement Problem: Given a set L of n lines in the plane, compute their arrangement which is a planar subdivision. Line Arrangements Problem: Given a set L of n lines
More information1 The range query problem
CS268: Geometric Algorithms Handout #12 Design and Analysis Original Handout #12 Stanford University Thursday, 19 May 1994 Original Lecture #12: Thursday, May 19, 1994 Topics: Range Searching with Partition
More informationLecture notes on the simplex method September We will present an algorithm to solve linear programs of the form. maximize.
Cornell University, Fall 2017 CS 6820: Algorithms Lecture notes on the simplex method September 2017 1 The Simplex Method We will present an algorithm to solve linear programs of the form maximize subject
More informationImproved Bounds for Intersecting Triangles and Halving Planes
Improved Bounds for Intersecting Triangles and Halving Planes David Eppstein Department of Information and Computer Science University of California, Irvine, CA 92717 Tech. Report 91-60 July 15, 1991 Abstract
More informationCrossing Families. Abstract
Crossing Families Boris Aronov 1, Paul Erdős 2, Wayne Goddard 3, Daniel J. Kleitman 3, Michael Klugerman 3, János Pach 2,4, Leonard J. Schulman 3 Abstract Given a set of points in the plane, a crossing
More informationMathematical and Algorithmic Foundations Linear Programming and Matchings
Adavnced Algorithms Lectures Mathematical and Algorithmic Foundations Linear Programming and Matchings Paul G. Spirakis Department of Computer Science University of Patras and Liverpool Paul G. Spirakis
More informationRobust Statistics and Arrangements David Eppstein
Robust Statistics and Arrangements David Eppstein Univ. of California, Irvine School of Information and Computer Science Two Worlds Robust Statistics Fit data points to model of data generation + errors
More informationJ Linear Programming Algorithms
Simplicibus itaque verbis gaudet Mathematica Veritas, cum etiam per se simplex sit Veritatis oratio. [And thus Mathematical Truth prefers simple words, because the language of Truth is itself simple.]
More informationLinear Programming in Small Dimensions
Linear Programming in Small Dimensions Lekcija 7 sergio.cabello@fmf.uni-lj.si FMF Univerza v Ljubljani Edited from slides by Antoine Vigneron Outline linear programming, motivation and definition one dimensional
More informationAn Efficient Algorithm for 2D Euclidean 2-Center with Outliers
An Efficient Algorithm for 2D Euclidean 2-Center with Outliers Pankaj K. Agarwal and Jeff M. Phillips Department of Computer Science, Duke University, Durham, NC 27708 Abstract. For a set P of n points
More informationOrthogonal Ham-Sandwich Theorem in R 3
Orthogonal Ham-Sandwich Theorem in R 3 Downloaded 11/24/17 to 37.44.201.8. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php Abstract The ham-sandwich theorem
More informationCMSC 754 Computational Geometry 1
CMSC 754 Computational Geometry 1 David M. Mount Department of Computer Science University of Maryland Fall 2005 1 Copyright, David M. Mount, 2005, Dept. of Computer Science, University of Maryland, College
More informationarxiv: v1 [cs.cc] 30 Jun 2017
On the Complexity of Polytopes in LI( Komei Fuuda May Szedlá July, 018 arxiv:170610114v1 [cscc] 30 Jun 017 Abstract In this paper we consider polytopes given by systems of n inequalities in d variables,
More informationG 6i try. On the Number of Minimal 1-Steiner Trees* Discrete Comput Geom 12:29-34 (1994)
Discrete Comput Geom 12:29-34 (1994) G 6i try 9 1994 Springer-Verlag New York Inc. On the Number of Minimal 1-Steiner Trees* B. Aronov, 1 M. Bern, 2 and D. Eppstein 3 Computer Science Department, Polytechnic
More informationOn the number of distinct directions of planes determined by n points in R 3
On the number of distinct directions of planes determined by n points in R 3 Rom Pinchasi August 27, 2007 Abstract We show that any set of n points in R 3, that is not contained in a plane, determines
More informationRestricted-Orientation Convexity in Higher-Dimensional Spaces
Restricted-Orientation Convexity in Higher-Dimensional Spaces ABSTRACT Eugene Fink Derick Wood University of Waterloo, Waterloo, Ont, Canada N2L3G1 {efink, dwood}@violetwaterlooedu A restricted-oriented
More informationRigidity of ball-polyhedra via truncated Voronoi and Delaunay complexes
!000111! NNNiiinnnttthhh IIInnnttteeerrrnnnaaatttiiiooonnnaaalll SSSyyymmmpppooosssiiiuuummm ooonnn VVVooorrrooonnnoooiii DDDiiiaaagggrrraaammmsss iiinnn SSSccciiieeennnccceee aaannnddd EEEnnngggiiinnneeeeeerrriiinnnggg
More informationApproximate Nearest Neighbor Problem: Improving Query Time CS468, 10/9/2006
Approximate Nearest Neighbor Problem: Improving Query Time CS468, 10/9/2006 Outline Reducing the constant from O ( ɛ d) to O ( ɛ (d 1)/2) in uery time Need to know ɛ ahead of time Preprocessing time and
More informationPreferred directions for resolving the non-uniqueness of Delaunay triangulations
Preferred directions for resolving the non-uniqueness of Delaunay triangulations Christopher Dyken and Michael S. Floater Abstract: This note proposes a simple rule to determine a unique triangulation
More informationImproved Results on Geometric Hitting Set Problems
Improved Results on Geometric Hitting Set Problems Nabil H. Mustafa nabil@lums.edu.pk Saurabh Ray saurabh@cs.uni-sb.de Abstract We consider the problem of computing minimum geometric hitting sets in which,
More informationMinimum-Link Watchman Tours
Minimum-Link Watchman Tours Esther M. Arkin Joseph S. B. Mitchell Christine D. Piatko Abstract We consider the problem of computing a watchman route in a polygon with holes. We show that the problem of
More informationHow to Cover Most of a Point Set with a V-Shape of Minimum Width
How to Cover Most of a Point Set with a V-Shape of Minimum Width Boris Aronov aronov@poly.edu John Iacono jiacono@poly.edu Özgür Özkan ozgurozkan@gmail.com Mark Yagnatinsky myag@cis.poly.edu Polytechnic
More informationINTRODUCTION TO THE HOMOLOGY GROUPS OF COMPLEXES
INTRODUCTION TO THE HOMOLOGY GROUPS OF COMPLEXES RACHEL CARANDANG Abstract. This paper provides an overview of the homology groups of a 2- dimensional complex. It then demonstrates a proof of the Invariance
More informationHowever, this is not always true! For example, this fails if both A and B are closed and unbounded (find an example).
98 CHAPTER 3. PROPERTIES OF CONVEX SETS: A GLIMPSE 3.2 Separation Theorems It seems intuitively rather obvious that if A and B are two nonempty disjoint convex sets in A 2, then there is a line, H, separating
More informationFor a set S of line segments, a separator can be found using duality. Under duality, the segments transform to double wedges, and a separator line tra
Separating and Shattering Long Line Segments? Alon Efrat School of Mathematical Sciences, Tel Aviv University, Tel-Aviv 69982, Israel. Email: alone@cs.tau.ac.il Otfried Schwarzkopf Dept of Computer Science,
More informationStabbers of line segments in the plane
Stabbers of line segments in the plane M. Claverol D. Garijo C. I. Grima A. Márquez C. Seara August 3, 2010 Abstract The problem of computing a representation of the stabbing lines of a set S of segments
More informationGeometric Red-Blue Set Cover for Unit Squares and Related Problems
Geometric Red-Blue Set Cover for Unit Squares and Related Problems Timothy M. Chan Nan Hu Abstract We study a geometric version of the Red-Blue Set Cover problem originally proposed by Carr, Doddi, Konjevod,
More informationAdvanced Operations Research Techniques IE316. Quiz 1 Review. Dr. Ted Ralphs
Advanced Operations Research Techniques IE316 Quiz 1 Review Dr. Ted Ralphs IE316 Quiz 1 Review 1 Reading for The Quiz Material covered in detail in lecture. 1.1, 1.4, 2.1-2.6, 3.1-3.3, 3.5 Background material
More informationFigure 2.1: An example of a convex set and a nonconvex one.
Convex Hulls 2.1 Definitions 2 Convexity is the key to understanding and simplifying geometry, and the convex hull plays a role in geometry akin to the sorted order for a collection of numbers. So what
More informationLinear Programming Duality and Algorithms
COMPSCI 330: Design and Analysis of Algorithms 4/5/2016 and 4/7/2016 Linear Programming Duality and Algorithms Lecturer: Debmalya Panigrahi Scribe: Tianqi Song 1 Overview In this lecture, we will cover
More informationEdge Guards for Polyhedra in Three-Space
Edge Guards for Polyhedra in Three-Space Javier Cano Csaba D. Tóth Jorge Urrutia Abstract It is shown that every polyhedron in R with m edges can be guarded with at most 27 2m The bound improves to 5 6
More informationA Reduction of Conway s Thrackle Conjecture
A Reduction of Conway s Thrackle Conjecture Wei Li, Karen Daniels, and Konstantin Rybnikov Department of Computer Science and Department of Mathematical Sciences University of Massachusetts, Lowell 01854
More informationOn s-intersecting Curves and Related Problems
On s-intersecting Curves and Related Problems Sarit Buzaglo, Ron Holzman, and Rom Pinchasi Mathematics Dept., Technion Israel Institute of Technology Haifa 32000, Israel sarahb@tx.technion.ac.il, holzman@tx.technion.ac.il,
More informationThe Simplex Algorithm
The Simplex Algorithm Uri Feige November 2011 1 The simplex algorithm The simplex algorithm was designed by Danzig in 1947. This write-up presents the main ideas involved. It is a slight update (mostly
More informationConvex hulls of spheres and convex hulls of convex polytopes lying on parallel hyperplanes
Convex hulls of spheres and convex hulls of convex polytopes lying on parallel hyperplanes Menelaos I. Karavelas joint work with Eleni Tzanaki University of Crete & FO.R.T.H. OrbiCG/ Workshop on Computational
More informationOn Covering a Graph Optimally with Induced Subgraphs
On Covering a Graph Optimally with Induced Subgraphs Shripad Thite April 1, 006 Abstract We consider the problem of covering a graph with a given number of induced subgraphs so that the maximum number
More informationCS 372: Computational Geometry Lecture 10 Linear Programming in Fixed Dimension
CS 372: Computational Geometry Lecture 10 Linear Programming in Fixed Dimension Antoine Vigneron King Abdullah University of Science and Technology November 7, 2012 Antoine Vigneron (KAUST) CS 372 Lecture
More informationOptimal detection of intersections between convex polyhedra
1 2 Optimal detection of intersections between convex polyhedra Luis Barba Stefan Langerman 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Abstract For a polyhedron P in R d, denote by P its combinatorial complexity,
More informationarxiv: v1 [math.co] 7 Dec 2018
SEQUENTIALLY EMBEDDABLE GRAPHS JACKSON AUTRY AND CHRISTOPHER O NEILL arxiv:1812.02904v1 [math.co] 7 Dec 2018 Abstract. We call a (not necessarily planar) embedding of a graph G in the plane sequential
More informationarxiv: v1 [cs.cg] 8 Jan 2018
Voronoi Diagrams for a Moderate-Sized Point-Set in a Simple Polygon Eunjin Oh Hee-Kap Ahn arxiv:1801.02292v1 [cs.cg] 8 Jan 2018 Abstract Given a set of sites in a simple polygon, a geodesic Voronoi diagram
More informationGeometric Red-Blue Set Cover for Unit Squares and Related Problems
Geometric Red-Blue Set Cover for Unit Squares and Related Problems Timothy M. Chan Nan Hu December 1, 2014 Abstract We study a geometric version of the Red-Blue Set Cover problem originally proposed by
More informationExact adaptive parallel algorithms for data depth problems. Vera Rosta Department of Mathematics and Statistics McGill University, Montreal
Exact adaptive parallel algorithms for data depth problems Vera Rosta Department of Mathematics and Statistics McGill University, Montreal joint work with Komei Fukuda School of Computer Science McGill
More informationAn Optimal Algorithm for the Euclidean Bottleneck Full Steiner Tree Problem
An Optimal Algorithm for the Euclidean Bottleneck Full Steiner Tree Problem Ahmad Biniaz Anil Maheshwari Michiel Smid September 30, 2013 Abstract Let P and S be two disjoint sets of n and m points in the
More informationOUTPUT-SENSITIVE ALGORITHMS FOR TUKEY DEPTH AND RELATED PROBLEMS
OUTPUT-SENSITIVE ALGORITHMS FOR TUKEY DEPTH AND RELATED PROBLEMS David Bremner Dan Chen John Iacono Stefan Langerman Pat Morin ABSTRACT. The Tukey depth (Tukey 1975) of a point p with respect to a finite
More informationACTUALLY DOING IT : an Introduction to Polyhedral Computation
ACTUALLY DOING IT : an Introduction to Polyhedral Computation Jesús A. De Loera Department of Mathematics Univ. of California, Davis http://www.math.ucdavis.edu/ deloera/ 1 What is a Convex Polytope? 2
More informationOn Computing the Centroid of the Vertices of an Arrangement and Related Problems
On Computing the Centroid of the Vertices of an Arrangement and Related Problems Deepak Ajwani, Saurabh Ray, Raimund Seidel, and Hans Raj Tiwary Max-Planck-Institut für Informatik, Saarbrücken, Germany
More informationApproximation Algorithms for Geometric Intersection Graphs
Approximation Algorithms for Geometric Intersection Graphs Subhas C. Nandy (nandysc@isical.ac.in) Advanced Computing and Microelectronics Unit Indian Statistical Institute Kolkata 700108, India. Outline
More informationSection Notes 5. Review of Linear Programming. Applied Math / Engineering Sciences 121. Week of October 15, 2017
Section Notes 5 Review of Linear Programming Applied Math / Engineering Sciences 121 Week of October 15, 2017 The following list of topics is an overview of the material that was covered in the lectures
More informationLecture 3: Art Gallery Problems and Polygon Triangulation
EECS 396/496: Computational Geometry Fall 2017 Lecture 3: Art Gallery Problems and Polygon Triangulation Lecturer: Huck Bennett In this lecture, we study the problem of guarding an art gallery (specified
More informationDiscrete Optimization. Lecture Notes 2
Discrete Optimization. Lecture Notes 2 Disjunctive Constraints Defining variables and formulating linear constraints can be straightforward or more sophisticated, depending on the problem structure. The
More informationMath 5593 Linear Programming Lecture Notes
Math 5593 Linear Programming Lecture Notes Unit II: Theory & Foundations (Convex Analysis) University of Colorado Denver, Fall 2013 Topics 1 Convex Sets 1 1.1 Basic Properties (Luenberger-Ye Appendix B.1).........................
More informationCrossing Numbers and Parameterized Complexity
Crossing Numbers and Parameterized Complexity MichaelJ.Pelsmajer 1, Marcus Schaefer 2, and Daniel Štefankovič3 1 Illinois Institute of Technology, Chicago, IL 60616, USA pelsmajer@iit.edu 2 DePaul University,
More informationClustering for Faster Network Simplex Pivots
Clustering for Faster Network Simplex Pivots David Eppstein Department of Information and Computer Science University of California, Irvine, CA 92717 Tech. Report 93-19 April 15, 1993 Abstract We show
More informationGeometric Streaming Algorithms with a Sorting Primitive (TR CS )
Geometric Streaming Algorithms with a Sorting Primitive (TR CS-2007-17) Eric Y. Chen School of Computer Science University of Waterloo Waterloo, ON N2L 3G1, Canada, y28chen@cs.uwaterloo.ca Abstract. We
More informationHYPERPLANE SEPARABILITY AND CONVEXITY OF PROBABILISTIC POINT SETS
HYPERPLANE SEPARABILITY AND CONVEXITY OF PROBABILISTIC POINT SETS Martin Fink, John Hershberger, Nirman Kumar, and Subhash Suri Abstract. We describe an O(n d )-time algorithm for computing the exact probability
More informationLine segment intersection. Family of intersection problems
CG Lecture 2 Line segment intersection Intersecting two line segments Line sweep algorithm Convex polygon intersection Boolean operations on polygons Subdivision overlay algorithm 1 Family of intersection
More informationThe Farthest Point Delaunay Triangulation Minimizes Angles
The Farthest Point Delaunay Triangulation Minimizes Angles David Eppstein Department of Information and Computer Science UC Irvine, CA 92717 November 20, 1990 Abstract We show that the planar dual to the
More information3 No-Wait Job Shops with Variable Processing Times
3 No-Wait Job Shops with Variable Processing Times In this chapter we assume that, on top of the classical no-wait job shop setting, we are given a set of processing times for each operation. We may select
More informationCS 473: Algorithms. Ruta Mehta. Spring University of Illinois, Urbana-Champaign. Ruta (UIUC) CS473 1 Spring / 36
CS 473: Algorithms Ruta Mehta University of Illinois, Urbana-Champaign Spring 2018 Ruta (UIUC) CS473 1 Spring 2018 1 / 36 CS 473: Algorithms, Spring 2018 LP Duality Lecture 20 April 3, 2018 Some of the
More informationLinear Data Structures for Fast Ray-Shooting amidst Convex Polyhedra
Linear Data Structures for Fast Ray-Shooting amidst Convex Polyhedra Haim Kaplan, Natan Rubin, and Micha Sharir School of Computer Science, Tel Aviv University, Tel Aviv, Israel {haimk,rubinnat,michas}@post.tau.ac.il
More information(for all > 0). In particular these bounds prove that none of the extent measures mentioned can change combinatorially more than O(n 2+ ) times. A quad
Submitted to the 1997 ACM Symposium on Computational Geometry Maintaining the Extent of a Moving Point Set Pankaj K. Agarwal Duke University Leonidas J. Guibas y Stanford University John Hershberger Mentor
More informationBichromatic Line Segment Intersection Counting in O(n log n) Time
Bichromatic Line Segment Intersection Counting in O(n log n) Time Timothy M. Chan Bryan T. Wilkinson Abstract We give an algorithm for bichromatic line segment intersection counting that runs in O(n log
More informationGeometric Unique Set Cover on Unit Disks and Unit Squares
CCCG 2016, Vancouver, British Columbia, August 3 5, 2016 Geometric Unique Set Cover on Unit Disks and Unit Squares Saeed Mehrabi Abstract We study the Unique Set Cover problem on unit disks and unit squares.
More informationExact Algorithms Lecture 7: FPT Hardness and the ETH
Exact Algorithms Lecture 7: FPT Hardness and the ETH February 12, 2016 Lecturer: Michael Lampis 1 Reminder: FPT algorithms Definition 1. A parameterized problem is a function from (χ, k) {0, 1} N to {0,
More informationThe Encoding Complexity of Network Coding
The Encoding Complexity of Network Coding Michael Langberg Alexander Sprintson Jehoshua Bruck California Institute of Technology Email: mikel,spalex,bruck @caltech.edu Abstract In the multicast network
More informationGeometry. Every Simplicial Polytope with at Most d + 4 Vertices Is a Quotient of a Neighborly Polytope. U. H. Kortenkamp. 1.
Discrete Comput Geom 18:455 462 (1997) Discrete & Computational Geometry 1997 Springer-Verlag New York Inc. Every Simplicial Polytope with at Most d + 4 Vertices Is a Quotient of a Neighborly Polytope
More informationOn the Relationships between Zero Forcing Numbers and Certain Graph Coverings
On the Relationships between Zero Forcing Numbers and Certain Graph Coverings Fatemeh Alinaghipour Taklimi, Shaun Fallat 1,, Karen Meagher 2 Department of Mathematics and Statistics, University of Regina,
More informationOn the Size of Higher-Dimensional Triangulations
Combinatorial and Computational Geometry MSRI Publications Volume 52, 2005 On the Size of Higher-Dimensional Triangulations PETER BRASS Abstract. I show that there are sets of n points in three dimensions,
More informationLecture 2 - Introduction to Polytopes
Lecture 2 - Introduction to Polytopes Optimization and Approximation - ENS M1 Nicolas Bousquet 1 Reminder of Linear Algebra definitions Let x 1,..., x m be points in R n and λ 1,..., λ m be real numbers.
More informationarxiv: v2 [cs.cg] 24 Jul 2011
Ice-Creams and Wedge Graphs Eyal Ackerman Tsachik Gelander Rom Pinchasi Abstract arxiv:116.855v2 [cs.cg] 24 Jul 211 What is the minimum angle α > such that given any set of α-directional antennas (that
More informationPoints covered an odd number of times by translates
Points covered an odd number of times by translates Rom Pinchasi August 5, 0 Abstract Let T be a fixed triangle and consider an odd number of translated copies of T in the plane. We show that the set of
More information6 Randomized rounding of semidefinite programs
6 Randomized rounding of semidefinite programs We now turn to a new tool which gives substantially improved performance guarantees for some problems We now show how nonlinear programming relaxations can
More informationSmallest Intersecting Circle for a Set of Polygons
Smallest Intersecting Circle for a Set of Polygons Peter Otfried Joachim Christian Marc Esther René Michiel Antoine Alexander 31st August 2005 1 Introduction Motivated by automated label placement of groups
More informationDiscussion of Location-Scale Depth by I. Mizera and C. H. Müller
Discussion of Location-Scale Depth by I. Mizera and C. H. Müller David Eppstein May 21, 2004 Mizera s previous work [13] introduced tangent depth, a powerful method for defining robust statistics. In this
More informationALGORITHMS FOR BALL HULLS AND BALL INTERSECTIONS IN NORMED PLANES
ALGORITHMS FOR BALL HULLS AND BALL INTERSECTIONS IN NORMED PLANES Pedro Martín and Horst Martini Abstract. Extending results of Hershberger and Suri for the Euclidean plane, we show that ball hulls and
More informationOutline. CS38 Introduction to Algorithms. Linear programming 5/21/2014. Linear programming. Lecture 15 May 20, 2014
5/2/24 Outline CS38 Introduction to Algorithms Lecture 5 May 2, 24 Linear programming simplex algorithm LP duality ellipsoid algorithm * slides from Kevin Wayne May 2, 24 CS38 Lecture 5 May 2, 24 CS38
More informationMinimum Bounding Boxes for Regular Cross-Polytopes
Minimum Bounding Boxes for Regular Cross-Polytopes Salman Shahid Michigan State University shahids1@cse.msu.edu Dr. Sakti Pramanik Michigan State University pramanik@cse.msu.edu Dr. Charles B. Owen Michigan
More informationTheoretical Computer Science
Theoretical Computer Science 408 (2008) 129 142 Contents lists available at ScienceDirect Theoretical Computer Science journal homepage: www.elsevier.com/locate/tcs Drawing colored graphs on colored points
More informationarxiv: v1 [math.co] 24 Aug 2009
SMOOTH FANO POLYTOPES ARISING FROM FINITE PARTIALLY ORDERED SETS arxiv:0908.3404v1 [math.co] 24 Aug 2009 TAKAYUKI HIBI AND AKIHIRO HIGASHITANI Abstract. Gorenstein Fano polytopes arising from finite partially
More information9 Bounds for the Knapsack Problem (March 6)
9 Bounds for the Knapsack Problem (March 6) In this lecture, I ll develop both upper and lower bounds in the linear decision tree model for the following version of the (NP-complete) Knapsack 1 problem:
More informationLecture 5: Duality Theory
Lecture 5: Duality Theory Rajat Mittal IIT Kanpur The objective of this lecture note will be to learn duality theory of linear programming. We are planning to answer following questions. What are hyperplane
More informationLecture 5: Simplicial Complex
Lecture 5: Simplicial Complex 2-Manifolds, Simplex and Simplicial Complex Scribed by: Lei Wang First part of this lecture finishes 2-Manifolds. Rest part of this lecture talks about simplicial complex.
More information66 III Complexes. R p (r) }.
66 III Complexes III.4 Alpha Complexes In this section, we use a radius constraint to introduce a family of subcomplexes of the Delaunay complex. These complexes are similar to the Čech complexes but differ
More informationRigid Ball-Polyhedra in Euclidean 3-Space
Discrete Comput Geom (2013) 49:189 199 DOI 10.1007/s00454-012-9480-y Rigid Ball-Polyhedra in Euclidean 3-Space Károly Bezdek Márton Naszódi Received: 15 September 2011 / Revised: 30 September 2012 / Accepted:
More informationChapter 2. Convex Hull. 2.1 Convexity. The following terminology should be familiar from linear algebra. Consider P R d. courses.
Chapter 2 Convex Hull 2.1 Convexity Consider P R d. courses. The following terminology should be familiar from linear algebra Linear hull. lin(p) := { q } q = λi p i 8 i : p i 2 P, λ i 2 R (smallest linear
More informationSome Advanced Topics in Linear Programming
Some Advanced Topics in Linear Programming Matthew J. Saltzman July 2, 995 Connections with Algebra and Geometry In this section, we will explore how some of the ideas in linear programming, duality theory,
More informationIce-Creams and Wedge Graphs
Ice-Creams and Wedge Graphs Eyal Ackerman Tsachik Gelander Rom Pinchasi Abstract What is the minimum angle α > such that given any set of α-directional antennas (that is, antennas each of which can communicate
More informationData Structures for Approximate Proximity and Range Searching
Data Structures for Approximate Proximity and Range Searching David M. Mount University of Maryland Joint work with: Sunil Arya (Hong Kong U. of Sci. and Tech) Charis Malamatos (Max Plank Inst.) 1 Introduction
More informationDetecting intersections between convex polyhedra
Detecting intersections between convex polyhedra Luis Barba Stefan Langerman Abstract For a polyhedron P in R d, denote by P its combinatorial complexity, i.e., the number of faces, edges and vertices
More informationIsometric Diamond Subgraphs
Isometric Diamond Subgraphs David Eppstein Computer Science Department, University of California, Irvine eppstein@uci.edu Abstract. We test in polynomial time whether a graph embeds in a distancepreserving
More informationLow-Dimensional Linear Programming with Violations
Low-Dimensional Linear Programming with Violations Timothy M. Chan October 21, 2004 Abstract Two decades ago, Megiddo and Dyer showed that linear programming in two and three dimensions (and subsequently
More informationAn Efficient Transformation for Klee s Measure Problem in the Streaming Model Abstract Given a stream of rectangles over a discrete space, we consider the problem of computing the total number of distinct
More informationCounting the Number of Eulerian Orientations
Counting the Number of Eulerian Orientations Zhenghui Wang March 16, 011 1 Introduction Consider an undirected Eulerian graph, a graph in which each vertex has even degree. An Eulerian orientation of the
More informationGeneralizing Ham Sandwich Cuts to Equitable Subdivisions
Discrete Comput Geom 24:605 622 (2000) DOI: 10.1007/s004540010065 Discrete & Computational Geometry 2000 Springer-Verlag New York Inc. Generalizing Ham Sandwich Cuts to Equitable Subdivisions S. Bespamyatnikh,
More informationImproved algorithms for constructing fault-tolerant spanners
Improved algorithms for constructing fault-tolerant spanners Christos Levcopoulos Giri Narasimhan Michiel Smid December 8, 2000 Abstract Let S be a set of n points in a metric space, and k a positive integer.
More informationAn Almost Space-Optimal Streaming Algorithm for Coresets in Fixed Dimensions
An Almost Space-Optimal Streaming Algorithm for Coresets in Fixed Dimensions Hamid Zarrabi-Zadeh Abstract We present a new streaming algorithm for maintaining an ε-kernel of a point set in R d using O((1/ε
More informationbe a polytope. has such a representation iff it contains the origin in its interior. For a generic, sort the inequalities so that
( Shelling (Bruggesser-Mani 1971) and Ranking Let be a polytope. has such a representation iff it contains the origin in its interior. For a generic, sort the inequalities so that. a ranking of vertices
More informationPebble Sets in Convex Polygons
2 1 Pebble Sets in Convex Polygons Kevin Iga, Randall Maddox June 15, 2005 Abstract Lukács and András posed the problem of showing the existence of a set of n 2 points in the interior of a convex n-gon
More informationFaster Construction of Planar Two-centers
Faster Construction of Planar Two-centers David Eppstein Abstract Improving on a recent breakthrough of Sharir, we find two minimum-radius circular disks covering a planar point set, in randomized expected
More information